Tool for automatic black-box parameter optimization released
Posted: Sun Jun 20, 2010 2:19 am
Hi,
A new version of my QLR tool is available. Download and screenshots are there:
http://remi.coulom.free.fr/QLR/
I first made a version that only works for Go (Win/Loss), but the only feedback I got from the computer-go mailing list was from chess programmers who tried to apply it to their programs. So now you can use the "chess" version of QLR. I thank Gian-Carlo in particular, for his enthusiastic beta testing.
If you have to optimize parameters, this tool should be much more efficient than manual optimization with bayeselo. Especially if you have a few correlated parameters. QLR should work very well up to a dozen parameters (not many more yet).
QLR is a bit complicated to set up, because you have to write a script for running games with values passed as parameters. In the future, I'll probably provide a ready-made script for xboard programs.
Rémi
A new version of my QLR tool is available. Download and screenshots are there:
http://remi.coulom.free.fr/QLR/
I first made a version that only works for Go (Win/Loss), but the only feedback I got from the computer-go mailing list was from chess programmers who tried to apply it to their programs. So now you can use the "chess" version of QLR. I thank Gian-Carlo in particular, for his enthusiastic beta testing.
If you have to optimize parameters, this tool should be much more efficient than manual optimization with bayeselo. Especially if you have a few correlated parameters. QLR should work very well up to a dozen parameters (not many more yet).
QLR is a bit complicated to set up, because you have to write a script for running games with values passed as parameters. In the future, I'll probably provide a ready-made script for xboard programs.
Rémi